Violin LabsViolin Labs

Ambitious AI ideas,
built fast.

Practical, end to end. You bring the problem. We ship the thing that solves it.

01The studio

Violin Labs is chill in temperament and fast in practice.

We take AI from a napkin sketch to a system running in production: agents, models, data pipelines, applied ML. We also build the foundations they run on: lakehouse architectures, governed data platforms, and the migrations that get you there. The whole path, not a slice of it.

Most engagements start the same way. You have a problem you can describe but not yet specify, and a deadline that makes a six-month discovery phase absurd. So we skip it. We'd rather hand you something running in week one than a proposal for something running in month six.

  • Agents
  • Models
  • Data pipelines
  • Applied ML
  • Lakehouse architecture
  • Data migrations
02How we work

Five moves, in the same order, every time.

01

Listen


What you do

Tell it to us in your own words. The mess, the constraint, the deadline.

What we do

Ask until we can say your problem back to you better than you said it.

02

Scope


What you do

Tell us what done looks like, and what it's worth.

What we do

Cut it to what matters. Say out loud what we are not building.

03

Promise


What you do

Hold us to a date.

What we do

Commit to a date and a shape. Short, honest, in writing.

04

Design


What you do

Push back on the diagram. It is cheapest to argue here.

What we do

Architect the system before we write it. The diagram has to survive reality.

05

Deliver


What you do

Use it. Break it. Tell us what's wrong.

What we do

Ship weekly, into production. Working software, not a demo.

Small bets, real prototypes, weekly drops. Scope stays honest and timelines stay short. The bar is whether it works in production, not whether it demos well. If it can be built in a week, it won't take a month.

03Agents in production

AI agents and applied ML running live in finance, legal, security, insurance, retail and media.

  • A Japanese telecom group

    Agent Factory

    A media spend optimisation agent answering business questions across 166+ tables, serving 400 production users in Japanese. Custom Python engine with deterministic lineage and audit trails.

  • A global consumer goods brand

    Trade Fund Manager

    Natural language queries over retail promotion budgets, via calibrated text-to-SQL with memory and access controls.

  • A private markets valuation platform

    Credit Agreement Parser

    Pulls loan terms out of 200+ page contracts into normalised schemas, using text segmentation and clause-level tracing.

  • A private markets valuation platform

    Financial Document Parsing

    Multimodal extraction with LayoutLM and RCNN, standardising fund and company financials. Cut manual effort by 90%.

  • A speech analytics company

    Call Analyzer

    Support call QA scoring across full call volume, combining Whisper ASR with LLM rubric grading.

  • An enterprise knowledge platform

    Legal Clause Finder

    Semantic contract search across 1,000+ clause types, built on contrastive embeddings with semantic caching.

  • A digital insurance carrier

    Vehicle Damage Detection

    Insurance claim photo classification with YOLO detection and continuous retraining on MLflow and Delta Lake.

  • An industrial electronics manufacturer

    Cybersecurity Sentinel

    Vulnerability detection in source code via fine-tuned Llama with automated evaluation, built for a national cyber challenge.

And some things we build for ourselves.

  • Our own tool

    CodePulse

    Semantic code context retrieval for Claude Code and Cursor, using PageRank-ranked graphs over MCP.

  • Our own tool

    Datalogist

    Maps raw data warehouses into AI-ready schemas through semantic inference and table clustering.

04Production on Databricks

Lakehouses, governed platforms, and the pipelines that feed them. Built to run, not to demo.

Medallion architectures, Unity Catalog governance, legacy migrations, production ML. If it runs on Databricks, we build it.

  • The country's leading fashion marketplace

    Analytics Platform Upgrade

    Apache Superset upgrade from 1.3.2 to 4.1.2 with zero dashboard loss and minutes-long rollback. Migrated 100 GB of MySQL metadata across major releases with sequential Alembic migrations and automated validation.

  • A specialty coffee chain

    Lakehouse Consolidation

    Unified fragmented Azure and AWS infrastructure into a governed Databricks platform with medallion architecture and Unity Catalog, serving 90+ retail locations.

  • A global consumer goods brand

    Agentic Analytics Platform

    Natural language queries over trade fund budgets using text-to-SQL agents, with access controls and audit trails built on Unity Catalog.

  • A private equity data platform

    Snowflake to Databricks Migration

    Enterprise data warehouse migration with phased cutover strategy, schema reconciliation, and zero-downtime transition to lakehouse architecture.

  • A travel booking platform

    MongoDB to Delta Lake Migration

    Production document database migration at scale, preserving schema flexibility while enabling SQL analytics and full data governance.

  • A digital insurance carrier

    ML Retraining Pipelines

    Continuous model retraining for vehicle damage detection, with MLflow experiment tracking and Delta Lake feature versioning.

  • A leading fintech company

    LLM Evaluation Pipeline

    Automated model comparison using LLM-as-a-judge scoring on Model Serving, with prompt optimization to determine version improvements.

  • A metropolitan transit authority

    Passenger Traffic Forecasting

    Time-series forecasting for rider volume prediction with MLOps workflow, scheduled retraining, and production deployment on Databricks.

05What we're drawn to

We do our best work on hard problems with clear shapes.

The clients we click with know what they want, or trust us to find out. The moments we look forward to are the first working prototype, and the diagram that survives contact with reality.

06What we won't do

The flip side, stated plainly.

  • No theatre.
  • No decks that outlive the product.
  • No scope that grows in the dark.
  • No “AI strategy” without a running system.
  • No pretending the model is the product when the data is the problem.
07Who we build for

A founder, operator, or team with a real AI problem and a real timeline.

You've thought about it enough to have an opinion, but not so much that you've ossified. You want a partner who codes, not a vendor who escalates. And you'll measure us on what ships.

If that's you

hello@theviolinlabs.com

Tell us the problem and the deadline. If we're the wrong shop for it, we'll say so.